Privacy-preserving mechanism for trajectory data publishing based on deep generative models

In order to overcome the problems such as poor trajectory data quality and insufficient privacy preservation in the trajectory data publishing, a privacy-preserving mechanism for trajectory data publishing based on deep generative models was proposed. Trajectory stop points were extracted by integrating multi-dimensional features such as time, distance, and speed, and the raw vehicle trajectories were segmented to reduce data redundancy and model training complexity. To effectively capture the spatio-temporal features in trajectory data, a trajectory synthesis model based on a generative adversarial network was designed by applying a long short-term memory network combined with a self-attention mechanism. The trajectory sequences were learned using a long short-term memory network and a self-attention mechanism, and then the model was trained with a generative adversarial network to generate high-quality synthetic trajectories. To further enhance the personalized privacy preservation of trajectories, a trajectory prediction model for users was designed by applying a bidirectional gated recurrent unit, and the model was trained with users' historical trajectory information. Through the learning and prediction mode, users' travel patterns were explored and analyzed from the training data to form personalized user trajectory prediction models. The synthetic trajectories were segmented and predicted by the trajectory prediction model. According to the prediction results, the trajectory segments requiring further enhanced privacy preservation were identified, with differential privacy noise added to improve privacy preservation, so as to obtain privacy-preserving trajectories for data publishing. Simulation results show that compared with existing methods, in the scenarios of taxi in Xi'an city and heavy truck trajectory data, the root-mean-square error reduces to 26 m. The JS divergences in spatial and temporal distributions reduce to 0.12 and 0.19, respectively, and the mutual information score reduces to 1.97. The proposed trajectory data preservation mechanism has been significantly improved in terms of trajectory availability and privacy preservation performance, demonstrating a good balance between privacy preservation and data utility.